Learning Meta-Distance for Sequences by Learning a Ground Metric via Virtual Sequence Regression.

Area of Science:

• Machine Learning
• Data Science
• Pattern Recognition

Background:

• Sequence data analysis requires understanding temporal alignments between vectors of varying lengths.
• Existing sequence distances often rely on a predefined ground metric, limiting adaptability.
• Learning adaptive meta-distances from complex sequences remains a significant challenge.

Purpose of the Study:

• To develop a method for learning sequence meta-distances by learning the underlying ground metric for vectors.
• To enable effective classification of multi-dimensional sequences based on learned distances.
• To address the challenge of inferring temporal alignments in variable-length sequences.

Main Methods:

• Formulating the ground metric as a learnable parameter within the meta-distance framework.
• Regressing sequences to virtual counterparts to ensure separation between classes.
• Developing iterative solutions for learning both Mahalanobis and deep neural network-induced metrics.
• Utilizing sequences with known ground metric-induced meta-distances as learning samples.

Main Results:

• The proposed method successfully learns ground metrics that induce effective meta-distances for sequence classification.
• Demonstrated ability to produce large meta-distances for inter-class sequences and small distances for intra-class sequences.
• Achieved effectiveness and efficiency across diverse sequence datasets.
• Validated the generalizability of the approach for various ground-metric-based sequence distances.

Conclusions:

• Learning the ground metric is a viable and effective strategy for learning sequence meta-distances.
• The proposed iterative learning framework is robust and applicable to different metric types.
• This work advances the field of sequence analysis and metric learning with practical implications.